There are thousands of bird species in the world, with numerous different and unique ones living in various areas. Developers Errol Joshua, Mahesh Nayak, Ajith K J, and Supriya Nickam wanted to build a simple device that would allow them to automatically recognize the feathered friends near them and do some simple tracking, such as knowing how often a particular bird makes its call. Their project uses a Nano 33 BLE Sense, along with its onboard microphone, to pick up sounds and make inferences about what they are in real-time. 

The team decided to train their tinyML model to detect four different species that are native to their area and then downloaded a sample dataset containing many sound files. After a bit of editing, they transferred the audio clips into Edge Impulse’s Studio and subsequently labeled each one. The Impulse consisted of a Mel-filter-bank energy (MFE) block that took the sounds and produced a spectrogram for each one. With these processed features, the model was able to achieve an impressive 95.9% accuracy. 

As seen in their demonstration video below, the current bird sound being played was picked up and identified accurately by the Nano 33 BLE Sense. And with some minor changes to how the model was trained, the accuracy can be increased even more. You can read about this project on its page.